Skip to Main content Skip to Navigation

Algorithmes décentralisés et asynchrones pour l'apprentissage statistique large échelle et application à l'indexation multimédia

Abstract : With the advent of the "data era", the amount of computational resources required by information processing systems has exploded, largely exceeding the technological evolutions of modern processors. Specifically, contemporary machine learning applications necessarily resort to massively distributed computation.Distributed algorithmics borrows most of its concepts from classical centralized and sequential algorithmics, where the system's behavior is defined as a sequence of instructions, executed one after the other. The importance of communication between computation units is generally neglected and pushed back to implementation details. Yet, as the number of units grows, the impact of local operations vanishes behind the emergent effects related to the large network of units. To preserve the desirable properties of centralized algorithmics such as stability, predictability and programmability, distributed computational paradigms must encompass this graph-theoretical dimension.This thesis proposes an algorithmic framework for large scale machine learning, which prevent two major drawbacks of classical methods, namely emph{centralization} and emph{synchronization}. We therefore introduce several new algorithms based on decentralized and asynchronous Gossip protocols, for solving clustering, density estimation, dimension reduction, classification and general convex optimization problems, while offering an appreciable speed-up on large networks with a very low communication cost. These practical advantages are mathematically supported by a theoretical convergence analysis. We finally illustrate the relevance of proposed methods on multimedia indexing applications and real image classification tasks.
Complete list of metadatas

Cited literature [257 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, April 25, 2018 - 2:22:06 PM
Last modification on : Thursday, March 5, 2020 - 4:25:49 PM
Long-term archiving on: : Monday, September 24, 2018 - 10:07:07 PM


Version validated by the jury (STAR)


  • HAL Id : tel-01778126, version 1



Jérôme Fellus. Algorithmes décentralisés et asynchrones pour l'apprentissage statistique large échelle et application à l'indexation multimédia. Algorithme et structure de données [cs.DS]. Université de Cergy Pontoise, 2017. Français. ⟨NNT : 2017CERG0899⟩. ⟨tel-01778126⟩



Record views


Files downloads